This uses the v2 768 model as the base model.

The dimensions work best as long as they're around 768, so a 768 by 768 works good but also a 512 by 768.
For high resolution steps: highres steps should be around 10 or 20, denoising strength should be at around 45 to 50, upscale shouldn't go higher then 1.5.


So the first question I'm sure is on your mind is why v2 768? in order to go forward, we need to go backwards, I can't run XL and I want to build a model I can use commercially.
The goal is to be able to produce higher quality artwork and have a lot of control over the model.

v2 768 doesn't have much content built into it already so it gives me a lot of control over the model and lets me ensure I don't have to fight the model to get rid of stuff that was already built into the base model.


The training data is the same I used from before, all the artwork I made over the years plus all the generations I've used from my own model, you're free to use it how you wish and make your own models with it.



If you want to train it yourself using my settings, you can get the training settings for free at my patreon and subscribestar pages:
https://www.patreon.com/posts/104114052?pr=true
https://subscribestar.adult/posts/1299771


If you just want to use my settings but don't know where to find the base model, you can find it here.
https://huggingface.co/stabilityai/stable-diffusion-2/tree/main